Source-linked AI summary
Locally computable error estimators for conforming approximations of interface problems cannot be robust
Yuwen Li
TL;DR
For interface problems, whether conforming-method a posteriori estimators can be both contrast-robust and locally computable remains open beyond restricted settings. The paper constructs checkerboard cross-point instances to show that extension locality obstructs this combination. Their energy errors separate by order M^1/4, ruling out contrast-independent reliability and efficiency for such estimators.
Problem
Whether conforming-method a posteriori estimators can be locally computable and robust to coefficient contrast remains unresolved for interface problems.
Source: §1 · pp. 1–2Method
The proof constructs two checkerboard instances with identical local data but different remote coefficient-branch continuations, using a piecewise-constant load invisible to conforming P1.
Source: §1 · p. 2, §3 · p. 6, §2.1 · p. 6Results
Order M^1/4: the two instances have energy-error scales separated by this factor, and no extension-local estimator can have both reliability and efficiency constants independent of contrast.
Source: §1 · p. 2, §4 · p. 12Takeaways & Limitations
The obstruction applies regardless of estimator structure, including residual, equilibrated-residual, and recovery-type estimators.
Source: §1 · p. 2, §2 · p. 3The theorem is framed for admissible conforming meshes and piecewise-constant coefficients taking values 1 and M, with a fixed extension radius.
Source: §2 · p. 3, §4 · p. 11
Abstract
from arXiv · showhide
We prove an impossibility result for extension-local a posteriori estimators for the conforming method for an elliptic interface problem at a checkerboard cross-point. We construct two diffusion problem instances on the same interface-fitted mesh that differ in the remote continuation of their coefficient-$M$ branches, yet every extension-local estimator satisfying efficiency must assign them the same value. They share a piecewise-constant load for which the finite element solution and the data oscillation both vanish. The ratio of their exact energy errors grows at least as the fourth root of the coefficient contrast $M$. Locality therefore forces the same estimator value for problems whose errors become increasingly different, yielding the central lower bound $C_{\rm eff}(M)C_{\rm rel}(M)\gtrsim M^{1/4}$, where $C_{\rm rel}(M)$ and $C_{\rm eff}(M)$ are the reliability and efficiency constants, respectively. Consequently, any locally computable error estimator, whether of residual, equilibrated, or recovery type, cannot be simultaneously reliable and efficient with contrast-independent constants, regardless of its algebraic form.
1. Introduction.
Prior robust a posteriori results for interface problems rely on quasi-monotonicity or concern nonconforming methods, leaving conforming methods an open question. The paper answers negatively for extension-local estimators by constructing a checkerboard cross-point obstruction.
1. Introduction.
- Residual estimators are robust for coefficient jumps under quasi-monotonicity, while previous non-quasi-monotone results do not exclude other robust, locally computable estimators.Earlier approximation results concern localization of best-approximation error, and a prior efficiency counterexample concerned a particular equilibrated estimator.Source: §1 · pp. 1–2
- CR and unified CR/interior-penalty DG estimators achieve robust a posteriori analysis without a quasi-monotonicity assumption, but do not resolve the conforming-method question.The modified CR estimator changes an edge-jump contribution near vertices where quasi-monotonicity fails.Source: §1 · p. 2
- At a checkerboard vertex with alternating coefficients 1, M, 1, M, two locally indistinguishable instances have energy-error scales separated by order M^1/4.This rules out robust extension-local estimation for the conforming P1 method, regardless of estimator structure.Source: §1 · p. 2
1.1. Problem setting.
- The problem setting uses conforming P1 finite elements on an interface-fitted mesh with checkerboard coefficients 1 in Q1 and Q3 and M in Q2 and Q4.The conforming solution is defined by the finite element variational equation.Source: §1.1 · pp. 2–3
2. Finite-range extension locality.
The section defines finite-range extension locality and the reliability and efficiency requirements for estimators. It then constructs a piecewise-constant load invisible to the conforming P1 method, with zero data oscillation.
2. Finite-range extension locality.
- An L-extension-local estimator assigns the same element indicator to problems that agree on the corresponding L-layer patch.The definition covers arbitrary estimator forms; residual, equilibrated-residual, and recovery estimators satisfy it.Source: §2 · pp. 3–4
- The analysis seeks reliability and efficiency constants bounded independently of contrast M; both bounds are required for jump-robustness.The estimator is considered through two-sided energy-error bounds, with a data-oscillation term subject to a consistency condition.Source: §2 · p. 4
- If the load and P1 solution vanish on an element’s L-layer patch, efficiency and locality force that element’s indicator to be zero.The proof compares the problem on that patch with the zero-load, zero-solution instance.Source: §2 · p. 4
2.1. A piecewise-constant invisible load.
- For every fixed locality radius L, the construction uses a mesh whose quality depends on L but not on M.The mesh keeps fixed macrotriangle subdivisions and inserts enough triangle layers to satisfy the stated neighborhood condition.Source: §2 · p. 4, §2.1 · pp. 5–6
- The piecewise-constant load has zero data oscillation and is invisible to the conforming P1 method, so the finite element solution is zero for every positive piecewise-constant coefficient.The load is supported in Q2 and Q4 and is independent of M; its total masses in the two coefficient-M branches are nonzero and opposite.Source: §2.1 · pp. 4, 6
- The two compared coefficient fields share the same domain and mesh, while the stated αA field has both coefficient-M quadrants extending from the cross-point to the Dirichlet boundary.The mesh is interface-fitted and includes the seven-triangle subdivision of each reflected macrotriangle.Source: Figure 2.1 · p. 5, §2.1 · p. 5, §3 · p. 6
3. Two problem instances.
The two problems share the same coefficient pattern near the checkerboard cross-point but differ in how the high-coefficient branches continue remotely. Both have vanishing finite-element approximations, while the section develops contrasting energy bounds for their exact solutions.
3. Two problem instances.
- Both finite-element approximations vanish although the coefficient fields differ in their remote continuation.The fields coincide throughout U1, where the αB branches terminate instead of reaching the physical boundary.Source: §3 · p. 6
3.1. Problem A.
- Lemma 3.1 provides an upper bound, in terms of M, on the exact-solution energy for Problem A.The proof constructs a flux supported in Q2 ∪ Q4, where αA = M.Source: §3.1 · pp. 7–8
3.2. Problem B.
- For Problem B, the section combines a positive load pairing with a controlled-energy test function to obtain a lower bound on the exact-solution energy.The test function's energy is bounded independently of M.Source: §3.2 · pp. 8–10
4. The no-go theorem.
The no-go theorem uses extension locality to make the two problems' indicators indistinguishable, despite their contrasting exact-solution energy bounds. It rules out reliability and efficiency constants that are both independent of coefficient contrast, even with any fixed logarithmic allowance.
4. The no-go theorem.
- The indicators agree because patches near the checkerboard lie where the coefficients coincide, while patches farther away have zero load and zero finite-element solution.The local-data argument establishes indistinguishability before the contrasting energy estimates are applied.Source: §4 · p. 11
- The theorem rules out reliability and efficiency constants that are both independent of coefficient contrast for extension-local estimators.Its assumptions include the reliability and efficiency bounds and a data-oscillation term satisfying the stated consistency condition.Source: §4 · pp. 11–12
- The obstruction also rules out estimators whose contrast dependence is bounded by any fixed logarithmic factor.The theorem's lower-bound growth dominates (1 + log M)^q for every fixed q ≥ 0.Source: §4 · p. 12